Infrastructure week. While the model launches keep coming, this week's biggest moves happened underneath the chatbots: Meta dropped the largest U.S. open-weight model ever released, DeepSeek open-sourced the inference engine behind its own production traffic, Nvidia's new Vera CPU scored its first AI cloud customer for agent hosting, OpenAI bought an AI biotech startup, and a startup that predicts real-world disruptions raised $325 million. Here is what happened, with every claim verified against primary sources.
Key Takeaways
Meta released Muse Spark Open (Oct 15): 816B parameters with 42B active per token, a 1M-token context window, and downloadable weights under a limited-commercial-use license, which Meta calls "the most capable U.S. open-weight model to date." (MacRumors)
DeepSeek open-sourced DeepInfer (Oct 20), the Apache 2.0 inference engine it says powers its own production services, supporting Nvidia SM100 GPUs and Huawei Ascend 950 NPUs. (TechTarget)
Nvidia's Vera CPU won AI cloud startup Modal as a customer for agent hosting; Modal's early tests say Vera doubled agents' completed work versus Intel, AMD, and Graviton chips (Modal's own numbers, not independently verified). (Reuters)
OpenAI acquired Orchestra Bio, a four-year-old AI biotech startup, folding its team into the Rosalind life-sciences unit. (TechCrunch)
ASI, a Boston startup building AI that forecasts physical-world disruptions, raised $325M at a $3B valuation and holds a $875M FAA contract. (Reuters)
1. Meta Releases Muse Spark Open: 816B Parameters of Open Weights
Just a week after launching its closed Muse Wonder models, Meta went the opposite direction. Muse Spark Open is an 816-billion-parameter open-weight model with 42 billion parameters active per token and a 1-million-token context window, reports MacRumors, citing Meta's October 15 announcement. Meta calls it "the most capable U.S. open-weight model to date."
It is natively multimodal (text, image, audio, and video input) and built for coding and long-horizon agentic work. The Apple partnership is the showpiece: Meta worked with Apple to run the model locally on two 512GB M5 Ultra Mac Studios through the MLX framework, with Meta's CEO thanking Apple's CEO in the announcement.
The weights are free to download from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry. The catch is the license, which allows only limited commercial use, so this is open weights with strings attached, not a permissive license like Apache 2.0. For builders, the download availability means real local experimentation, including on high-end Apple silicon.
If you want quick video breakdowns of trending open-source AI projects as they ship, browse the repo signals for 30-second overviews.
2. DeepSeek Open-Sources DeepInfer, Its Production Inference Engine
DeepSeek published DeepInfer on GitHub under the Apache 2.0 license on October 20, the inference framework it says powers its own production services, reports TechTarget. This is not a model release, it is the serving stack behind DeepSeek's famously cheap API.
The scope is deliberately narrow for now: DeepInfer supports two chip families, Nvidia SM100 GPUs and Huawei Ascend 950 NPUs, and currently runs DeepSeek-V4.1-Flash in offline mode only. The techniques are the interesting part: custom kernels, MoE expert parallelism, speculative decoding with draft tokens, and structured output via regex, JSON schema, and EBNF. It is built primarily on vLLM, borrowing ideas from SGLang and TensorRT-LLM.
DeepSeek's own production figures for V4.1-Flash are striking: 155 trillion tokens per day per cluster, a 96% input cache hit rate, about 400 milliseconds average time-to-first-token, and 230 tokens per second (up to 300 per second for coding), all noted as DeepSeek's figures and not independently verified. The same TechTarget piece notes that the U.S. banned federal agencies from using DeepSeek earlier this week, with China calling the move politicization.
3. Nvidia's Vera CPU Wins Modal for AI Agent Hosting
Nvidia's new Vera central processor has its first public design win in AI agent hosting: Modal, a fast-growing AI cloud startup, will run Vera chips for hosting AI agents, moving off Intel and AMD entirely for its agent sandboxes, Reuters reports.
The headline number comes from Modal's own early tests: Vera doubled the work completed by AI agents compared with Intel Granite Rapids, AMD Zen 5, and AWS Graviton chips (Modal's numbers, not independently verified). That framing matters, because the workload that wins is agents, not classic cloud VMs.
The deal's context is just as loud. Modal reportedly raised $750 million last week led by Accel at a $15.75 billion valuation, with usage tripling and revenue up more than 15 times year over year, targeting $1 billion in sales over the next year. Nvidia expects at least $20 billion in Vera chip sales this fiscal year in a CPU market it values at roughly $200 billion, and OpenAI, Anthropic, Oracle, and Perplexity also plan to use Vera. The CPU, not the GPU, is now the contested frontier for the agent economy.
4. OpenAI Acquires Orchestra Bio for Its Life-Sciences Push
OpenAI acquired Orchestra Bio, a four-year-old AI biotech software startup, on October 20, with terms undisclosed, reports TechCrunch. Orchestra had reportedly raised around $20 million as of 2025 from Andreessen Horowitz, Atria, and 8VC.
The interesting detail is what Orchestra actually builds: AI agents that manage drug-discovery operations minutiae, like finance tracking and clinical-trial scheduling. Its team and technology join OpenAI's Rosalind life-sciences unit, home of the GPT-Rosalind model and the Rosalind Workbench. Named Rosalind customers include Moderna and Novo Nordisk.
This is OpenAI buying operational AI agents rather than another model lab: the value is in boring, high-stakes workflow automation for biotech, where scheduling and tracking mistakes are expensive. It is the same agentic theme running through this whole week's news, applied to drug discovery.
5. ASI Raises $325M to Predict Disruptions Before They Happen
ASI, a Boston startup building AI systems that forecast physical-world disruptions, raised a $325 million Series C at a $3 billion valuation, led by Spark Capital with participation from Citadel Securities affiliates, Meritech Capital, and Craft Ventures, Reuters reports.
The pitch is a deliberate contrast with the LLM world. Founder and CEO Phillip Buckendorf: "While the broader tech ecosystem spent hundreds of billions teaching AI to manipulate words and generate text, ASI is deeply focused on AI systems that can manage and control physical assets." ASI's software simulates and monitors systems to forecast disruptions across air traffic, power grids, and military operations.
The contracts back up the story. ASI holds a five-year, $99 million U.S. Army contract awarded in October and a twelve-year, $875 million FAA contract awarded in June for air traffic management. That is $974 million in public-sector commitments against a $3 billion valuation, a rare ratio for an AI startup at this stage.
How the Week's Stories Stack Up
Aspect | Story | Key Number | Openness |
|---|---|---|---|
Biggest open weights | Muse Spark Open (Meta) | 816B params, 42B active | Weights public, limited commercial license |
Inference stack | DeepInfer (DeepSeek) | 155T tokens/day/cluster (own figures) | Apache 2.0 |
Agent silicon | Vera CPU (Nvidia + Modal) | 2x agent work completed (Modal's tests) | Hardware sale |
Biotech agents | Orchestra Bio (OpenAI) | ~$20M raised pre-acquisition | Acquired, closed |
Physical-world AI | ASI Series C | $325M at $3B valuation | Private company |
The thread across all five: agents are the workload that matters. Meta built an open-weight model for long-horizon agentic work, DeepSeek open-sourced the serving stack that makes agents cheap, Nvidia's Vera is being sold on agent throughput, OpenAI bought agents for drug discovery, and ASI raised on agents for the physical world.
Frequently Asked Questions
Is Meta's Muse Spark Open really open source?
Not exactly. The weights are free to download, which makes it open-weight, but the license allows only limited commercial use. That is less permissive than a true open license like Apache 2.0, so read the terms before building a commercial product on it.
What does DeepSeek's DeepInfer actually do?
DeepInfer is an inference engine, the software that serves model responses at scale, not a model itself. DeepSeek says it powers its own production services, and it now supports Nvidia SM100 GPUs and Huawei Ascend 950 NPUs with techniques like MoE expert parallelism and speculative decoding.
Why does Nvidia's Vera CPU matter for AI agents?
Agents run as sandboxes that need fast CPUs to coordinate tool calls, memory, and retries, not just GPU flops. Modal's early tests claim Vera doubled agents' completed work versus Intel, AMD, and Graviton chips, and Nvidia expects $20 billion in Vera sales this fiscal year.
What is OpenAI's Rosalind unit?
Rosalind is OpenAI's life-sciences unit, home to the GPT-Rosalind model and the Rosalind Workbench. The Orchestra Bio acquisition adds AI agents that automate drug-discovery operations like finance tracking and clinical-trial scheduling. Named customers include Moderna and Novo Nordisk.
Who is ASI and why did they raise $325M?
ASI is a Boston startup building AI that forecasts and manages disruptions across physical systems like air traffic, power grids, and military operations. Backed by $99 million in Army contracts and an $875 million FAA contract, it raised $325 million at a $3 billion valuation led by Spark Capital.
Conclusion
This week was about what happens below the surface of AI products. Meta proved open-weight flagships can be a U.S. strategy, DeepSeek gave the serving stack away for free, Nvidia's CPU found its agent-era customer, and two deals (OpenAI buying Orchestra Bio, ASI's $325 million raise) bet big that the next AI frontier is agents operating on real work: drug trials, air traffic, power grids. Next week's OpenAI announcements will show whether the model layer has an answer.
Read this article in Bahasa Indonesia: Berita AI: 14-21 Oktober 2026.